Mao zheng
Technical Expert of Tencent Hunyuan, Head of Hunyuan Application Algorithms
Graduated with a Master’s degree from Harbin Institute of Technology, currently serving as a Senior Technical Expert at Tencent Hunyuan, leading the application algorithm team. Responsible for driving the deployment of Hunyuan large models across Tencent, supporting applications in advertising, social networking, entertainment, education, and customer service, with a focus on optimizing performance at the application layer. His research interests include complex reasoning with large models, multi-turn dialogue, translation, RAG, and agent technologies. He has published dozens of papers at top conferences such as ACL, AAAI, EMNLP, COLING, and CVPR. Under his leadership, the team has achieved first place in multiple tracks of the WMT translation competition.
Topic
Experience Sharing on Optimization of Tencent Hunyuan Translation Models
Tencent Hunyuan Open-Source International Translation Competition Champion Model Hunyuan-MT-7B (7 billion parameters) supports translation among 33 languages and 5 Chinese dialects. Additionally, Tencent released the industry’s first open-source translation ensemble model, Hunyuan-MT-Chimera-7B, which reached the top of the Hugging Face trending list within a week of release. In the ACL WMT2025 competition, Hunyuan-MT-7B achieved first place in 30 out of 31 evaluated languages. Despite its parameter limitations, it outperformed larger models, and Flores200 evaluation shows it leading other models of similar scale. The model has been applied in Tencent Meeting, WeCom (Tencent Enterprise WeChat), QQ Browser, Tencent Customer Service, and other business scenarios. This talk focuses on the model’s training and optimization techniques as well as its practical deployment experience. Outline: Challenges Facing Machine Translation in the Era of Large Models Building Large-Model Machine Translation Capabilities SHY training paradigm design Key technical challenges in the CPT stage Key technical challenges in the SFT stage Key technical challenges in the RL stage Challenges Encountered in Business Applications Optimization of domain-specific terminology translation Improvement of translation performance for low-resource languages
